MLOps

Twiga ships a production layer that closes the loop between model development and real-world deployment. The four modules below form a coherent MLOps stack that mirrors Twiga’s config-driven, Pydantic-first design principles.

Modules

• MLOps Architecture
  • Overview
  • Package structure
  • Forecaster ↔ MLOps integration
  • Component architecture
  • Data contracts
  • Data flow - complete production cycle
  • Configuration - TwigaSettings
  • Design principles
• Experiment Tracking
  • Overview
  • Quick start
  • API reference
  • Logging helpers
  • ExperimentEngine tracking
  • Using a remote tracking server
  • Prefect integration
• Checkpoints
  • Saving
  • Loading
  • In the MLOps pipeline
  • API reference
• Model Serving
  • Overview
  • Quick start
  • Endpoints
  • API reference
  • ModelLoader
  • Deployment tips
• Production Monitoring
  • Overview
  • Quick start
  • Report types
  • HTML reports
  • API reference
  • Integration with the FastAPI app
  • Integration with Prefect
• Pipeline Orchestration
  • Overview
  • training_flow
  • retraining_flow
  • Pipeline schemas
  • Tasks
  • Scheduling with Prefect deployments
  • End-to-end pipeline

Installation

Install the full MLOps stack with a single extra:

pip install twiga[mlops]

This pulls in MLflow, FastAPI, uvicorn, Evidently, and Prefect.

At a glance

Module	Class / function	Responsibility
twiga.tracking	TwigaTracker	MLflow experiment tracking
twiga.forecaster	on_save_checkpoint	Persist model + pipeline with versioned manifest
twiga.forecaster	on_load_checkpoint	Restore model + pipeline from checkpoint
twiga.serve	create_app()	FastAPI REST API factory
twiga.serve	ModelLoader	Checkpoint to forecaster loader
twiga.serve	ForecastMonitor	Evidently drift & performance reports
twiga.pipeline	training_flow	Prefect training orchestration
twiga.pipeline	retraining_flow	Drift-triggered retraining

twiga.mlops core API

A small, Streamlit-free layer for workspace orchestration, dataset transforms, storage, capture-data access, and the training model registry. Any script or service can import it directly — from twiga.mlops import workspace needs no Streamlit runtime.

Module	Class / function	Responsibility
twiga.mlops.workspace	create_workspace, load_workspace_data, LoadedWorkspace	Create / load / reconfigure workspaces; returns data, no session state
twiga.mlops.data	parse_dataset, split_raw_frame, build_configs_from_setup	Pure dataset parsing, splitting, and config construction
twiga.mlops.catalog	Catalog, WorkspaceRow	SQLite registry of workspaces (thread-safe writes)
twiga.mlops.storage	LocalFsStorage, local_tracking_uri_for	Per-workspace filesystem layout (dataset, mlruns, checkpoints, reports)
twiga.mlops.monitoring	read_predictions, read_actuals, predictions_vs_actuals, daily_capture_counts, run_batch	Read and join captured production traffic; batch drift/performance
twiga.mlops.mlflow_query	list_runs, latest_run	Convenience wrappers over the MLflow tracking store
twiga.mlops.training	get_registry, build_model_config, ModelEntry	Lazy model registry + per-run config construction
twiga.mlops.paths	MLOPS_ROOT, STATE_DB_PATH, WORKSPACES_DIR	Catalog DB / workspace root locations (override via TWIGA_MLOPS_ROOT)

Note

The multi-page Streamlit dashboard that drives this layer lives in examples/mlops/ and is not shipped in the installed wheel — it is a demo, like the notebooks under examples/tutorials/. Run it from a checkout:

pip install "twiga[mlops]" streamlit
cd examples/mlops
streamlit run app.py